Machine learning, particularly deep learning, has been spectacularly successful in the recent past. Human-level performance has been achieved in tasks such as object recognition and speech recognition. However, these successes are achieved for highly constrained domains that requires minimal amounts of reasoning and planning, and use large amounts of labeled training data.

This work aims to incorporate knowledge about model structures, constraints, prior knowledge, as well as inference and planning algorithms into the machine learning methods in order to extend their reach on problems that require reasoning and planning. In particular, we aim develop methods that combine deep learning methods with probabilistic graphical models, in order to exploit the strengths of both methods, enabling models to be trained with less training data, and scaling up machine learning to work on more complex problems.